Track record

Backtest: the learned discriminator (logistic model, trained on the 58 resolved drops) has its winner-vs-dud call compared to the real resale outcome under cross-validation — 68.8% directional accuracy vs a 52% baseline. The in-sample fit on the same 58 drops reads higher; it is not the number the desk quotes. Live: every forecast we publish is logged with a timestamp and graded when the drop resolves; each resolved outcome is fed back into the training set, so the model sharpens itself. We publish the misses too — an honest record is the whole point.